Skip to content
New issue

Have a question about this project? Sign up for a free GitHub account to open an issue and contact its maintainers and the community.

By clicking “Sign up for GitHub”, you agree to our terms of service and privacy statement. We’ll occasionally send you account related emails.

Already on GitHub? Sign in to your account

Learning Type-Aware Embeddings for Fashion Compatibility #360

Open
chullhwan-song opened this issue Apr 10, 2020 · 1 comment
Open

Learning Type-Aware Embeddings for Fashion Compatibility #360

chullhwan-song opened this issue Apr 10, 2020 · 1 comment

Comments

@chullhwan-song
Copy link
Owner

https://arxiv.org/abs/1803.09196

@chullhwan-song
Copy link
Owner Author

chullhwan-song commented Apr 11, 2020

abstract

  • Outfits = 많은 다른 타입의 구성 > top, bottom, shoes
  • similarity vs compatibility
  • jointly learns notions of item similarity and compatibility

개념

image
image

구조

image

  • image > 18-layer Deep Residual Network
  • text > HGLMM Fisher vector encoding [17] of word2vec

Respecting Type in Embedding

  • parametern theta를 학습 시키는 embedding feature > 학습 network
    image
  • compatibility pairt image
  • image taxonomy(분류) 타입
  • 이 조건에서, triplet를 구성하기 위해서, u, v 의 type의 개념을 삽입 > similairty 가 아니고 compatibility임을 인식
  • image triplet로 구성 > 이 anchor 와 pos, neg는 그 타입(카테고리)이 다르다. 다만, 이때, pos==neg은 같은 타입
    image
    • anchor image image
    • positive image image
    • negative image image
  • 정리하자면, anchor image 와 카테고리(items in outfit)가 다르지만, (anchor image, positive image) 는 compatible 하고, (anchor image, negative image) 는 non-compatible란 의미
  • 이렇게 구성한 triplet를 이용한 loss를 다음과 같이
    image

Constraints on the learned

  • 부가적으로 similairty 와 mulit-modal visual semantic embedding 방법을 추가
  • similairty
    image
  • visual semantic embedding > text와 image간의 embedding기법 > 이전 리뷰들 참고 예 - DeViSE: A Deep Visual-Semantic Embedding Model #1
    image
  • 최종 loss
    image

Experiment

image

  • 그냥 compatibility+ triplet loss가 가장 효과가 있고, visual semantic embedding & similarity는 그닥 도움이 안되는 것 같다.
  • compatibility > 근데 패션에 대해 몰라서 그런지..잘 어울리는지 모르겠다.
    image
    image
    image
  • fill-in-the-blank task
    image
  • 예 결과
    image
    image
    image

Repository owner deleted a comment from peternara Apr 11, 2020
Sign up for free to join this conversation on GitHub. Already have an account? Sign in to comment
Projects
None yet
Development

No branches or pull requests

1 participant